SUMMARY

Rolling Element Bearings (REB) are critical components of a wide range of rotating machines. Identifying and preventing their faults is critical for safe and efficient equipment operation. A variety of condition monitoring techniques have been developed that gather large amounts of data using acoustic or vibration transducers. Further information about the health of an REB can be extracted via time domain trend analysis, and amplitude modulation technics. The frequency domain-specific peaks corresponding to the defects can also be identified directly from the spectrum. Such approaches either provide little insight into the type of defect, are sensitive to noise, and require substantial post-processing. Complicating current fault diagnostic approaches are the ever-increasing size of datasets from different types of sensors that yield non-homogeneous databases and more challenging to execute prognostics for large-scale condition-based maintenance. These difficulties are addressable via approaches that leverage recent developments on microprocessors and system on chip (SoC) enabling more processing power at the sensor level, unloading the cloud from non-used or low information density data. The proposed research addresses these limitations by presenting a new approach for bearing defect detection using a SoC network to perform a wavelet transform calculation. The wavelet transforms enable an improved time-frequency representation and is less sensitive to noise than other classical methods; however, its analysis requires more complex processing techniques that must be executed at the edge (sensor) to limit the need for cloud computing of the results and large-scale data transmission to the cloud. To enable near real-time processing of the data, the BeagleBone AI SoC is employed, the wavelet transforms, and the defect classification are achieved at the edge. The contributions of this work are as follows: first, the real-time data acquisition driver for the SoC is developed. Second, the machine learning algorithm for improving the wavelet transform and the defect identification is implemented. Third federated learning in a network of SoC is formulated and implemented. Finally, the new approach is benchmarked to current approaches in terms of detection accuracy, and sensitivity to defect and was proven to obtain between 80 and 90 percent accuracy depending on the dataset. https://teams.microsoft.com/l/meetup-join/19%3ameeting_MWZjZWRmMzEtZDY5OS00ZjJjLTkzNGItYTgyMTkwMzU0YjNi%40thread.v2/0?context=%7b%22Tid%22%3a%22482198bb-ae7b-4b25-8b7a-6d7f32faa083%22%2c%22Oid%22%3a%2260831a82-f78a-4f78-bab0-3801b0e1b5a0%22%7d